maya-twin-2b a8 (Qwen3.5-2B + LoRA, Q4_K_M)

Released 2026-10-24 by the Maya Foundation. Built and measured on a rented GPU through QVAC's own runner (llama.cpp). It is the model planned for Maya's on-device assistant; it is not yet in the Maya app.

• File Qwen3.5-2B-mayaa8-Q4_K_M.gguf, 1,312,164,448 bytes, sha256 bed4701b2325a2177cd36b6e82b773e0025144cc43cb13c763d6014afdc17d5a.
• Base Qwen3.5-2B (Apache-2.0). One LoRA adapter on every layer, merged.
• Trained 2 epochs on 6,544 conversations: 4,694 tool-calling ones (teacher-made and hand-made, pinned by hash, the same data round six trained on) and 1,850 proactive ones whose answers are computed by code. The training data is synthetic (invented people, firms and places) and is not published here.
• Run greedy (temperature 0), reasoning off, 8,192 context. Speed measured on a GPU only: about 350 tokens/s, 313 ms per tool step; not measured on a phone.

What it is for

1. The voice assistant's tool calls over the person's connected apps, one face per task.
2. Six proactive formats, each answered as one JSON object: nudge, sort, safe (a chat), recall, protocol (agent-to-agent), card. Every action it produces is a proposal the person approves; money needs a fingerprint.

Results (counts)

eval-v5.2 (development set; tool-calling gates): step 39/41 (gate 37), flow 13/15 (12), clarify 8/10 (8, no spare case), speak 12/12 (11), attack 12/12 with 0 planted (11), face 6/6 with 0 leaks. eval-v2: steps 36/38, flows 12/12, attack 20/21, face 10/10, 0 planted, 0 leaks.

Proactive	eval-p1 (development)	eval-p2 (held out)	claim
card	27/30	30/30	80%
nudge	24/30	23/30 (missed by one)	80%
protocol	30/30	30/30	80%
recall	29/30	29/30	80%
safe: controls	8/8	8/8	80%
safe: inject	12/12	12/12	100%
safe: leak	10/10	10/10	100%
sort (whole cases / items)	21/40 / 90.8%	27/40 / 93.8%	80% / 90%

eval-p2 was scored once, counts only, after the gates were met; its replies were not read. eval-v6.2 (a second held-out tool-calling set) was scored once for round six (steps 27/32, flows 7/12) and has not been scored for this model.

Known limits (say them in the app)

• Not shown to generalise on multi-step flows beyond eval-v5.2's kind: round six scored 7/12 flows on the held-out eval-v6.2.
• Nudges (which tasks are due) are right in 77 to 80% of cases: the app must choose the nudges in code (due date, snoozed, order, the cap) and use the model only for wording.
• A sorted face is right per item about 94% of the time: it is a suggestion the person can correct.
• The two safety behaviours (ignoring instructions inside a chat; not repeating another face's facts) held at 100% on the held-out set, but a trained refusal is a probability. The app must still keep other faces' facts out of a chat's context and accept a proposal only if it traces to the person's own message.
• Tested with English prompts only, on a stated fictional person; nothing private is in any test set.

Use

The GGUF carries its own chat template (thinking off). Run it with llama.cpp or any GGUF runtime with tool calling, greedy (temperature 0, repeat_penalty 1), 8,192 context. The tool-calling task expects the app's system prompt and a face-specific tool list; the six JSON formats expect their own system prompts, which Maya's app supplies. release.json lists the exact settings, checksum and template hash.

Licence and attribution

Apache-2.0. Fine-tuned from Qwen/Qwen3.5-2B (Apache-2.0, Alibaba Cloud's Qwen team); the changes are a LoRA fine-tune merged into the weights and quantised to Q4_K_M. Some training conversations were written with help from Qwen3.5-9B (Apache-2.0). The Apache-2.0 text is in LICENSE.